About large model based AI systems, because why not?

The public availablity of ChatGPT, DALLE2 and related or adjunctive systems like Bing has set the news cycle on fire and every company from A to Z is trying to incorporate large model artificial intelligence into their products regardless of whether these things are well suited for the task. If you want to write a sales pitch, brief article, or essay; ChatGPT can help and will even provide references. It will gladly argue any position you assign it along with supporting references in a clear and concise manner. If you want to generate some artwork or a photo like the one above, DALLE2 is your tool. Meanwhile, Bing's implementation is very Microsoft as in embrace and extend like most of their historical activity in creating software solutions. It has live access to the Internet and can search for any information to support even a very complex query requiring multiple steps to look up and summarize findings.

This all sounds quite lovely and marketing, sales, and research folk are tapping into these systems in droves. There is an inherent issue though. The data these things have been trained on is publicly available Internet content. This content is quite diverse in quality, accuracy, and bias. The outputs of these systems can vary a great deal in a similar manner. With DALLE2, the outputs can often be nightmarish, mysoginistic, or copyright infringements. With ChatGPT, the extensive disclaimers pushed onto it can't entirely hide or interdict the systems ability to produce biased or inaccurate content.

With Bing's live Internet accessible extension to ChatGPT, this gets worse. As an example, I asked a fairly complex question with a numerical answer that I already knew, but one that would require both research of publicly available Internet content in multiple locations as well as solid math skills to get us there. It asked me two relevant questions needed to perform the work (despite that one of those was answered in my question) and after I provided those (one again), poof! It gave me an answer that was wrong. I could tell where it went astray, pointed out the issue and asked it to revise it's answer. It acknowledged the error and researched the necessary adjusting data and recalculated. Poof! It was wrong again. Again I could see what it was missing, pointed out the problem and asked for another revision. It again acknowledged the error, did the necessary research and calculations and poof! It finally spat out the right number.

It had to do it this way, because it doesn't understand anything about the topic or how things work in the real world. People will absolutely anthropomorphize this thing as that one guy who was fired for sharing chats and claiming it was sentient. I'll agree that it's extremely smart , capable, and can simulate sentience very well. It really isn't though. Some have compared it to a predictive cell phone text software on steroids. It's more than that, but not by much.

This simulated intelligence causes two very real problems. The first is like Tesla's fully self driving autopilot, people will trust it when it is not actually capable of the task or earning the trust it has been given. Imagine training one as an FAA air traffic controller in an environment with no substantial terrain or weather and perfect data. It would do just fine. Add terrain variations, weather, spotty RADAR returns and poor communications and things will get ugly very quickly. The ability to handle edge cases well just isn't a possibility with something like this that actually doesn't understand the environment it operates in. Automated weapons systems are already getting these types of models and there is much to fear about a very real impending skynet styled scenario.

On the flip-side, these models can absolutely reduce man-hours and costs for researching and creating content similar to what they were trained on. Will they be properly restricted to these purposes? Absolutely not. Bing is just the tip of the public adoption iceberg as firms race to throw similar AI systems into every product they make. In my own company, we are exploring using ChatGPT and DALLE2 to help produce some marketing content. Marketing and graphic editing professionals are doing that work not to replace any humans, but to supplement them so they can be more effective and efficient. That represents a best use case as far as I can see. Other than that, getting a good answer to your queries is a complete crap-shoot, but I'm sure they'll be using something similar to write school text books soon. Be careful and Good luck!